Reliable agents
Recover cleanly when tools, providers, and state fail.
Co-founder & CTO · Systems builder
Infrastructure for AI that has to work.
I design AI systems that stay reliable under load, predictable under failure, and efficient under constraint.
Current focus2026—
Cost-aware systems
Control spend across a workflow, not one request at a time.
Reproducible testing
Treat agent behavior as software that can be verified.
Selected experience
Building production-ready agent systems.
Led Kubernetes migration at company scale.
2018 Amazon Alexa Prize winner.
Selected work / 01
Systems built for the messy part.
The interesting problems begin after the demo: traffic spikes, provider failures, runaway costs, and workflows that need to resume cleanly.
Open source · Agent testing01
Open source · LLM infrastructure02
Operating principles / 02
Make constraints part of the architecture.
Expect failure
Design recovery, fallback, and observability before the happy path ships.
Route with context
Optimize whole workflows—not isolated requests—for quality, latency, and cost.
Test the system
Treat agents like applications: reproducible, idempotent, and accountable.
Field notes / 03
Thinking in public.
010203
LLM Routers Are Not Enough
Why per-request routing misses the point—and what workflow-level cost control looks like.
Reproducible Testing Reveals the Hidden Risk in Autonomous Agents: Idempotency
Why autonomous agents need deterministic testing, and how idempotency failures surface in production.
Agents Should Be Tested Like Applications, Not Evaluated Like Models
The case for treating agent systems as software—with integration tests, not just evals.
Toolbox / 04
Production AI infrastructureFailure-aware orchestrationMulti-provider reliabilityBudget-constrained routingKubernetesDistributed systems